Catastrophic Forgetting

The fundamental tension between a network's plasticity (ability to learn new information) and its stability (ability to retain old information). When fine-tuning on a narrow domain like clinical text, the gradient updates aggressively shift weights to minimize the new loss, overwriting the configurations that encoded general language understanding. This is not a bug but an inherent trade-off in connectionist models.

The mechanistic cause where weight updates in the deep layers of a transformer destroy previously learned feature representations. For example, a model fine-tuned exclusively on radiology reports may lose its ability to perform general sentiment analysis or basic arithmetic. This occurs because the high-level feature space is repurposed entirely for the narrow target domain, collapsing the manifold of general knowledge.

A specific manifestation where learning Task B degrades performance on Task A without revisiting Task A's data. In clinical NLP, a model sequentially fine-tuned on MIMIC-III notes and then on PubMed abstracts will likely forget the clinical entity recognition patterns learned from MIMIC-III. This is due to destructive interference in the shared parameter space.

A foundational algorithmic defense that identifies parameters critical to previous tasks and penalizes their modification. EWC calculates the Fisher Information Matrix to estimate the importance of each weight. During new task training, a quadratic penalty term is added to the loss function, anchoring crucial weights near their old values and forcing learning into less critical parameters.

A data-centric strategy that interleaves samples from previous tasks into the current training batch. By maintaining a small memory buffer of representative general-domain examples, the model is constantly reminded of its original capabilities. This prevents the loss landscape from shifting exclusively toward the new domain, effectively rehearsing old knowledge to maintain synaptic stability.

Techniques like Low-Rank Adaptation (LoRA) inherently resist catastrophic forgetting by freezing the vast majority of pre-trained weights. Instead of modifying the original weight matrix W, LoRA learns a low-rank decomposition ΔW = BA. The original knowledge remains intact in the frozen backbone, while the small, trainable adapters capture domain-specific nuances without disturbing the base representations.

A family of adaptation methods that freeze the vast majority of pre-trained weights and inject only a tiny fraction of new, trainable parameters. By limiting the degrees of freedom during optimization, PEFT acts as a strong regularizer that preserves the original model's general knowledge while learning a new domain.

• LoRA injects low-rank matrices into attention layers.
• Adapters insert small bottleneck layers between transformer blocks.
• Prefix Tuning prepends learnable virtual tokens to the input.

A rehearsal-based strategy that interleaves samples from the original pre-training data with the new domain-specific data during fine-tuning. By maintaining a memory buffer of prior examples, the model continuously revisits its old knowledge, preventing the gradient updates from overwriting established representations.

• Maximal Interference Retrieval selects replay samples that are most likely to be forgotten.
• Generative Replay uses a separate generative model to synthesize pseudo-samples of past data.

An architectural approach that freezes the original model entirely and learns a new task by instantiating a new, separate neural network column. Lateral connections from the frozen base model to the new column allow the system to leverage previously learned features without modifying them. While computationally expensive, this guarantees zero forgetting and is useful when absolute preservation of a validated base model is non-negotiable.

A sequential training strategy that bridges the distribution gap before task-specific fine-tuning. Instead of jumping directly from general text to a narrow medical task, the model first undergoes continued unsupervised pre-training on a large, unlabeled domain corpus like PubMed abstracts or MIMIC-III clinical notes. This allows the model to internalize the statistical properties of medical language broadly, reducing the shock of the final specialization step.

A technique where a smaller student model is trained to mimic the output distribution of a larger, frozen teacher model on both the old and new tasks. The teacher provides soft targets that encode the rich similarity structure of its general knowledge. By distilling this into the student alongside the new task objective, the student learns to perform the new task without catastrophically diverging from the teacher's original representational space.